The signal behind the sentence
Financial news rarely announces its meaning in simple positive or negative language. A company raises its outlook, a competitor enters its market, or an acquisition falls through. Understanding the significance requires identifying what happened, who was involved, and how that event changes the reader's view of a business. A document-level sentiment label leaves much of that information behind.
SENTiVENT grew from Gilles Jacobs's doctoral research at Ghent University, with collaboration across language technology and financial research. The work addressed a practical gap: detailed, annotated economic news that could support supervised event extraction and fine-grained sentiment analysis. That academic foundation continues to inform Exverbai's approach to applied language technology.
Design the evidence before training the model
The project brought data collection, annotation design, model development, and error analysis into one research process. A business-news corpus and an economic event typology established a shared description of the task. Annotation guidelines made difficult distinctions explicit, including the difference between an actual event, a possibility, and a negated statement.
The result connects event mentions with participants and attributes rather than reducing articles to isolated keywords. Expert annotation and an adjudicated reference set made it possible to assess system output against a consistent target. Compatibility with established event-extraction conventions also made the resource useful beyond a single experimental model.
From text classification to structured understanding
The research investigated economic event detection and the implicit sentiment carried by factual descriptions. This distinction matters whenever language supplies evidence for a downstream decision. An article may contain several events with different implications for different companies; the useful representation preserves those relationships instead of averaging them away.
For a business application, the same reasoning changes how an NLP system should be scoped. Define the information needed by the analyst, preserve its connection to source text, and evaluate each extraction step. Named entities, event categories, participant roles, and sentiment then become inspectable components of an information product.
A foundation that can be reused
The project produced peer-reviewed research and a publicly available dataset. Subsequent independent work makes that foundation accessible to contemporary ML and LLM evaluation workflows. The research contribution and the later engineering work have different contexts, but share an emphasis on explicit task definitions and reproducible evidence.
SENTiVENT illustrates the value of domain-specific AI engineering: turn a loosely stated objective into a concrete information task, build reference data that reflects its difficulty, and make errors visible. Those habits transfer to financial intelligence, document analysis, and knowledge systems where a convincing answer must also be a defensible one.
